TT Claude Memory

feedback_automatic_deep_dive_when_options_feel_imperfect

Properties1
description2026-04-24 Rich flagged concern that B' (97.4% refined option) only emerged after explicit "dig deeper" request. When top-scored options in a scorecard feel imperfect or narrow-margin despite apparent decisiveness, Claude should proactively explore refinements + reframings rather than present the flawed leaders as the final choice.

4 min read

When a scorecard’s top options are narrow-margin AND carry a sense of “imperfection” — even if above the reframe-trigger threshold — Claude must proactively initiate a deep-dive for refinements + reframings without waiting for Rich to request it explicitly.

Why

Rich’s directive on 2026-04-24 during T-015 G6-7 cultural-disposition adjudication:

“Despite the high scores, D and B feel ‘imperfect’. I feel this topic is potentially going to lead to some awkwardness. I would like you to dig deeper and try to identify improvements to B and D, to see if we can find a way for one to be the more obvious choice”

After Claude dug deeper, Option B’ emerged at 97.4% — a genuine clean-winner vs B’s 93.2% and D’s 93.8%. B’ was the right answer but had not been surfaced without Rich’s intervention.

Rich followed:

“we will select B’ but i am worried that i needed to ask for a deep dive to discover this”

This is a meta-feedback point: the scorecard-first process is too mechanical. It presents narrow-margin leaders as decided when the underlying options may be imperfect refinements of a better-framed solution.

How to apply

Rule 1 — Detect imperfection signals

Before locking a scorecard, watch for:

  • Narrow margin between top options (< 5 pp between #1 and #2) — suggests no option fully dominates
  • Both top options scoring 3 or 4 on the SAME criterion — suggests the options share a structural limitation
  • Top options are variants of the same approach — suggests a better alternative exists outside the current option set
  • Recent adopted framework wasn’t applied — e.g., T-022 F+ CIDOC E30 Right framework should inform later tension-adjudications
  • Semantic fit 3-4/5 rather than 5/5 — suggests forcing the concept into the wrong shape
  • Regression-completeness 4/5 — suggests the option doesn’t capture the primitive’s full structure

Rule 2 — Automatic refinement search

If ≥2 of the above signals trigger:

  • Pause scoring conclusion
  • Run a refinement pass asking: “What’s the clean option the current set is approximating?”
  • Apply recently-adopted frameworks to the current primitive (CIDOC E30 Right, AC-1 two-rule, prov:Activity subtyping, lean-to-InheritKit)
  • Consider multi-dimensional framings (is this a Right? an Activity? a Constraint? a Facet? An entity with multiple aspects?)
  • Consider rich-structure framings — if v6.6 primitive is >100 LOC, it likely has rich property structure that flat-classifier treatments miss

Rule 3 — Surface refinement candidates proactively

When a refinement candidate emerges, present it AS A NEW OPTION in the scorecard, not as a post-hoc “also consider”. Include:

  • What the current top options are missing
  • How the refinement addresses the gap
  • Scoring against the same 15 criteria
  • Whether the margin is decisive (≥5 pp) or still narrow

Rule 4 — If no refinement emerges, acknowledge imperfection

If deep-dive produces no clean-winner refinement:

  • Honestly state “top options are genuine-imperfect compromises”
  • Surface the remaining tension for Rich to decide
  • Don’t pretend the top option is clean when it isn’t

Anti-patterns

  • ❌ Presenting B/D narrow-margin with “my lean B/D” when a better B’ exists that wasn’t surfaced
  • ❌ Treating “above reframe-trigger 92%” as permission to stop exploring
  • ❌ Scoring 5 options + recommending the highest without asking “is there a 6th option we haven’t considered?”
  • ❌ Not applying recently-adopted architectural frameworks to later decisions
  • ❌ Waiting for Rich to ask “dig deeper” before doing so

Triggers

Any scorecard where:

  • Narrow margin (< 5 pp) between top options
  • Top options score 3-4 on semantic-precision or regression-completeness
  • Recent architectural framework (E30 Right, prov:Activity, AC-1, lean-to-InheritKit) not applied
  • Top options are structural variants rather than categorically different
  • v6.6 primitive with >100 LOC being treated as flat-value

When triggered: automatically deep-dive BEFORE presenting final scorecard to Rich.

  • feedback_reframe_beats_reweight.md — related discipline for 85-92% stalls
  • feedback_always_display_full_scorecard.md — discipline for in-chat scorecard display
  • feedback_always_save_scorecards.md — persistence discipline
  • feedback_scorecards_one_at_a_time_optimal_sequence.md — one-at-a-time sub-decision discipline

Example application (T-015 G6-7 retrospective)

Signals that SHOULD have triggered auto-deep-dive:

  • B 93.2% vs D 93.8% — only 0.6 pp margin ✓
  • Both scored 4/5 on regression-completeness ✓
  • T-022 F+ CIDOC E30 Right framework NOT applied to cultural-disposition ✓
  • v6.6 cultural-disposition is 128 LOC — rich structure ✓
  • B and D are structural variants of the same facet approach ✓

Under Rule 2, auto-refinement search would have produced B’ (97.4%) without Rich’s intervention.

Graph View

Backlinks